Shape Representation and Recognition from Multiscale Curvature

摘要

We present a technique for shape representation and the recognition of objects based on multiscale curvature information. It provides a single framework for both the decomposition and recognition of both planar curves as well as surfaces in three-dimensional space. The decomposition operation simultaneously performs data interpolation, data smoothing, and segmentation. The unification of these three stages results in a smoothing operation that is coupled with the primitives to be used in description. Each of the minimization operators, in addition to having a curvature tuning, also has a different spatial sensitivity function. As a result, the different possible descriptions capture information at multiple spatial scales. This allows a single region of an object to be described in more than one way, when appropriate. The practicality of the ensuing representation is demonstrated by the recognition of planar curves. A matching strategy based on dynamic programming is used. The results illustrate the manner in which a continuous spectrum of similar objects can be defined, ranging from those that are very similar to a target to those that are very different from it.